Detecting people in dense crowds

Abstract

We propose a scheme to detect individuals in any image frame of a video sequence showing densely crowded scenes against cluttered backgrounds. The method uses only spatial information, and in an initial pass through the image a trained Viola–Jones-type local detector is used to locate individuals in the densely crowded scene. This yields a large number of false alarms. Hence, in a second step, we seek to reduce the false alarms, and propose two methods for this. In the first, color information from the initially detected windows is passed to a classifier to reduce the false alarms. This classifier consists of a cascade of boosted classifiers with Haar-like features as input and is trained with color information from local windows. In the second method, a weak perspective model of an uncalibrated camera is used to further reduce the false alarm rate while maintaining the detection rate. This is based on the size and locations of the detections in the image frame, without the use of any 3D world information. Results are presented in the form of receiver operating characteristic curves. For instance, at a 79.0% detection accuracy, the false alarm rate is 20.3%.